A conditional random field is a discriminative statistical modelling method that is used when the class labels for different inputs are not independent. For example, in image segmentation, the class label for the pixel depends on the label of its neighboring pixels also. If (X1, X2...Xn) denotes the observations (or features) and (Y1, Y2, Y3...Yn) denotes the label, a discrete classifier such as SVM, predicts a value of Yi for given Xi considering single input at a time. Here the assumption is that the different inputs are independent.
However, when there is a high degree of correlation between the input data points and the labels, such classifiers tend to underperform. A CRF on the other hand provides prediction for the data point Yi considering features Xi as well as the labels of all other nodes which are dependent of node i. In CRF, data points are arranged as graph consisting of a set of nodes V, and edges between nodes E. An edge between node i and node j denotes labels of data point i and j are dependent. CRF is used in sequential data processing use cases such as POS tagging in NLP and image segmentation in computer vision.
Sunil kumar's session on Conditional Random Fields (CRFs)
The below image is an example for Semantic Segmentation.
- Segmentation --> Partition an image into multiple segments
- Semantic Segmentation --> Segments are meaningful regions as perceived by human
Source: Picture taken from “TextonBoost” for Image Understanding: Multiclass Object recognition and Segmentation by jointly modelling Texture, Layout and Context”. Jamie Shotton et. al. IJCV 2009
Click here to learn more about AI, ML, DL related concepts.